Psych 156A/ Ling 150: Psychology of Language Learning

1
Psych 156A/ Ling 150Psychology of Language
Learning

• Lecture 9
• Structure

2
Announcements

• Reminder Office hours for Lisa this week on
  Thursday from 330 - 530pm, not today.
• Pick up midterms
• Homework 2 due today
• Review questions posted for phrases
• Homework 3 posted (due 2/24/09)

3
About Language Structure

• Sentences are not just strings of words.
• The girl danced with the goblin king.

4
About Language Structure

• Sentences are not just strings of words.
• Words cluster into larger units called phrases,
  based on their grammatical category.

Noun (N) girl, goblin, dream, laughter,
Determiner (Det) a, the, an, these,
Adjective (Adj) lovely, stinky, purple,
Verb (V) laugh, dance, see, defeat, Adverb
(Adv) lazily, well, rather, Preposition (P)
with, on, around, towards,
5
About Language Structure

• Sentences are not just strings of words.
• Words cluster into larger units called phrases,
  based on their grammatical category.

Det
Det
N
N
V
P
Adj
The girl danced with the Goblin King.
6
About Language Structure

• Sentences are not just strings of words.
• Words cluster into larger units called phrases,
  based on their grammatical category.

Det
Det
N
N
V
P
Adj
The girl danced with the Goblin King.
Noun Phrases (NP)
7
About Language Structure

• Sentences are not just strings of words.
• Words cluster into larger units called phrases,
  based on their grammatical category.

Det
Det
N
N
V
P
Adj
The girl danced with the Goblin King.
Noun Phrases (NP)
Can be replaced with pronouns like he, she,
it,
8
About Language Structure

• Sentences are not just strings of words.
• Words cluster into larger units called phrases,
  based on their grammatical category.

Det
Det
N
N
V
P
Adj
She danced with him.
Noun Phrases (NP)
Can be replaced with pronouns like he, she,
it,
9
About Language Structure

• Sentences are not just strings of words.
• Words cluster into larger units called phrases,
  based on their grammatical category.

Det
Det
N
N
V
P
Adj
The girl danced with the Goblin King.
Preposition Phrases (PP)
10
About Language Structure

• Sentences are not just strings of words.
• Words cluster into larger units called phrases,
  based on their grammatical category.

Det
N
Det
N
V
P
Adj
The girl danced with the Goblin King.
Preposition Phrases (PP)
Can be replaced with words like here and there
11
About Language Structure

• Sentences are not just strings of words.
• Words cluster into larger units called phrases,
  based on their grammatical category.

Det
N
Det
N
V
P
Adj
The girl danced there.
Preposition Phrases (PP)
Can be replaced with words like here and there
12
About Language Structure

• Sentences are not just strings of words.
• Words cluster into larger units called phrases,
  based on their grammatical category.

Det
Det
N
N
V
P
Adj
The girl danced with the Goblin King.
Verb Phrases (VP)
13
About Language Structure

• Sentences are not just strings of words.
• Words cluster into larger units called phrases,
  based on their grammatical category.

Det
N
Det
N
V
P
Adj
The girl danced with the Goblin King.
Verb Phrases (VP)
Can be replaced with words like do so and did
so
14
About Language Structure

• Sentences are not just strings of words.
• Words cluster into larger units called phrases,
  based on their grammatical category.

Det
N
Det
N
V
P
Adj
The girl did so.
Verb Phrases (VP)
Can be replaced with words like do so and did
so
15
About Language Structure

• Sentences are not just strings of words.
• Words cluster into larger units called phrases,
  based on their grammatical category.

Det
Det
N
N
V
P
Adj
The girl danced with the goblin king.
Verb Phrases (VP)
Preposition Phrases (PP)
Noun Phrases (NP)
16
About Language Structure

• Sentences are not just strings of words.

Sentence
Another way to represent this phrase structure
VP
PP
NP
NP
Det
Det
N
N
V
P
Adj
The girl danced with the goblin king.
17
Computational Problem

• How do children figure out which words belong
  together (as phrases) and which words dont?

Det
Det
N
N
V
P
Adj
The girl danced with the goblin king.
Det
Det
N
N
V
P
Adj
The girl danced with the goblin king.
18
Learning Phrases

• One way weve seen that children can learn things
  is by tracking the statistical information
  available.
• Saffran, Aslin, Newport (1996)
• Transitional Probability is something
  8-month-olds can track

Prob(stlebe) lt Prob(castle) Prob(stlebe) lt
Prob(beyond)
to the castle beyond the goblin city
Posit a word boundary at the minimum of the
transitional probabilities between syllables
19
Learning Phrases

• One way weve seen that children can learn things
  is by tracking the statistical information
  available.
• Thompson Newport (2007)
• Transitional Probability to divide words into
  phrases?

the girl and the dwarf
Posit a phrase boundary where the transitional
probability is low between words?
20
A look at real language properties in action with
transitional probabilities
Example Optional phrases
A B C D E F
The goblin easily steals the child.
21
A look at real language properties in action with
transitional probabilities
Example Optional phrases
A B C D E F
The goblin easily steals the child.
If the child only ever sees this order of
categories, theres no way to know how the words
break up into phrases using transitional
probabilities. Why? TrProb(AB) TrProb(BC)
TrProb(CD) TrProb(DE) TrProb(EF) 1
ABCDEF
22
A look at real language properties in action with
transitional probabilities
Example Optional phrases
A B C D E F
The goblin easily steals the child.
But suppose C is an optional word/phrase. (easily
is an adverb that can be left out)
ABCDEF
ABDEF
Data without C sometimes will appear.
The goblin steals the child.
23
A look at real language properties in action with
transitional probabilities
Example Optional phrases
A B C D E F
The goblin easily steals the child.
With the optional phrase left out, TrProb(BC) is
less than 1 since sometimes B is followed by D
instead of always being followed by C. A
transitional probability learner later
encountering ABCDEF might posit a phrase boundary
between B and C because Tr(AB) and TrProb(CD) are
still 1.
ABCDEF
ABDEF
The goblin steals the child.
24
A look at real language properties in action with
transitional probabilities
Example Optional phrases
A B C D E F
The goblin easily steals the child.
Conclusion AB is a unit, CDEF is a unit. the
goblin ( NP) easily steals the child ( VP)
ABCDEF
ABDEF
The goblin steals the child.
25
A look at real language properties in action with
transitional probabilities
Example Optional phrases
A B C D E F
The goblin easily steals the child.
For ABDEF, Tr(AB) and Tr(DE) 1, but TrProb(BD)
lt 1. So, a transitional probability learner will
posit a boundary between B and D. Conclusion AB
is a unit, DEF is a unit. the goblin ( NP)
steals the child ( VP)
ABCDEF
ABDEF
The goblin steals the child.
26
Artificial Language Experiments
Adults (not children) listened to data from an
artificial language for 20 minutes on multiple
days Assumption Adults who are learning an
artificial language will behave like children who
are learning their first language since the
adults have no prior experience with the
artificial just as children have no prior
experience with their first language Is this a
good assumption to make?
27
Adults in Artificial Language Experiments
Children in First Language?
Maybe yes, if childrens brains are able to
track the same information as adults brains.
Then, the fact that adults can learn phrases from
transitional probabilities means children should
also be able to learn phrases from transitional
probabilities. Maybe no, if there are other
factors that could interfere, such as adults
having more cognitive resources to process
information or using their native language
experience to help them learn something about the
artificial language. Then, just because adults
succeed doesnt mean children will also succeed.
28
Artificial Language Similar To Real Language?
Properties of the artificial language that are
similar to real language properties optional
phrases (the goblin chased a chicken in the
castle) PP is optional in the
sentence repeated phrases (NP Verb
NP PP) More than one NP is used in
the sentence moved phrases (In the castle the
goblin chased a chicken) PP is moved to the
front of the sentence
29
Artificial Language Experiments
Baseline pattern ABCDEF
real language parallel
A B C D E F
The goblin easily steals the child.
Artificial Language Phrases AB CD EF
30
How do we tell if learning happened?
Baseline assessment Can subjects actually
realize all these nonsense words belong to 6
distinct categories? Can they categorize?
kof hox jes sot fal ker is the same as daz
neb tid zor rud sib
31
How do we tell if learning happened?
Baseline assessment Can subjects actually
realize all these nonsense words belong to 6
distinct categories? Can they categorize?
See if they can tell the difference between
the correct order they were exposed to (ABCDEF)
and some other pattern they never heard
(ABCDCF)
kof hox jes sot fal ker is the same as daz
neb tid zor rud sib
kof hox jes sot fal ker is right kof hox jes
sot rel ker is wrong
32
How do we tell if learning happened?
Phrase learning assessment If they can
categorize, do they learn what the phrases are
(AB, CD, EF)?
Example test between AB and non-phrase
BC Sample test item - which one do they think
belongs together?
kof hox vs. hox jes
33
Learning a language with optional phrases
Baseline pattern ABCDEF
Other patterns heard (phrases AB CD EF
missing) CDEF, ABEF, ABCD kof hox jes
sot fal ker rel zor taf nav mer neb rud
sib daz lev tid lum
Control subjects Control language (remove one
adjacent pair at a time) Additional control
patterns heard BCDE, ABCF, ADEF
34
Learning a language with optional phrases
Transitional Probabilities in the Optional Phrase
language and the Control language are different.
The Optional Phrase language has lower
probability across phrase boundaries than within
phrases. The control language has the same
probability no matter what.
35
Learning a language with repeated phrases
Baseline pattern ABCDEF
Other patterns heard (phrases AB CD EF
repeated) ABCDEFAB, ABCDEFCD, ABCDEFEF
kof hox jes sot fal ker kof hox rel zor taf
nav daz neb mer neb jes zor rud sib tid sot
daz lev tid lum fal nav taf ker
Control subjects Control language (repeat one
adjacent pair at a time) Additional control
patterns heard ABCDEFBC, ABCDEFDE, ABCDEFFA
36
Learning a language with repeated phrases
Transitional Probabilities in the Repeated Phrase
language and the Control language are different.
The Repeated Phrase language has lower
probability across phrase boundaries than within
phrases. The control language has almost the same
probability no matter what.
37
Learning a language with moved phrases
Baseline pattern ABCDEF
Other patterns heard (phrases AB CD EF
moved) ABCDEF, ABEFCD, CDABEF, CDEFAB,
EFABCD, EFCDAB Example strings heard
kof hox jes sot fal ker daz neb taf nav
rel zor
Control subjects Control language (move one
adjacent pair at a time) Additional control
patterns heard BCAFDE, AFDEBC, DEAFBC, DEBCAF
38
Learning a language with moved phrases
Transitional Probabilities in the Moved Phrase
language and the Control language are different.
The Moved Phrase language has lower probability
across phrase boundaries than within phrases. The
control language has the same probability no
matter what.
39
Artificial Language Learning Categorization,
Day 1
Generally above chance performance (50), control
group performing about the same or a little worse
than test groups.
40
Artificial Language Learning Categorization,
Day 5
General improvement, though test groups still a
little better than control groups. Still,
subjects generally capable of categorization.
Mean correct for all subjects is significantly
above chance (which would be 50)
41
Artificial Language Learning Phrases, Day 1
In each case, even after only 20 minutes of
exposure (day 1), test subjects are better than
control subjects for each of the languages with
optional, repeated, or moved phrases.
42
Artificial Language Learning Phrases, Day 5
After 5 days of exposure (100 minutes), the
difference between control subjects and test
subjects becomes apparent.
control?? Human tendency towards binary groupings
43
Artificial Language Learning Phrases, Day 5
After 5 days of exposure (100 minutes), the
difference between control subjects and test
subjects becomes apparent.
control?? Human tendency towards binary groupings
Some properties seem easier to pick up on than
others (repeated and movement language subjects
are much better than control subjects).
44
Artificial Language Learning Phrases, Day 5
After 5 days of exposure (100 minutes), the
difference between control subjects and test
subjects becomes apparent.
control?? Human tendency towards binary groupings
Interestingly, control subjects in the optional
phrase condition actually did really well - this
is unexpected since the transitional
probabilities were uninformative.
45
Learning a language with optional phrases,
repeated phrases, and moved phrases
Baseline pattern ABCDEF
Other patterns heard (phrases AB CD EF
moved) CDEF, ABEF, ABCD, ABCDEFAB, ABCDEFCD,
ABCDEFEF, ABCDEF, ABEFCD, CDABEF, CDEFAB, EFABCD,
EFCDAB
Transitional Probabilities in the All-combined
language and the Control language are different.
The All-combined language has lower probability
across phrase boundaries than within phrases. The
control language probabilities are more uniform,
though they do vary.
46
Predictions for all-combined condition?

• One idea Harder
• Why? There are many more patterns that are
  acceptable for the artificial language. Even if
  transitional probability is informative, its a
  lot of information to track because there are so
  many patterns that are acceptable and even more
  potential patterns that are unacceptable.
• Prediction Test subjects dont do much better
  than control subjects.
• Second idea The same, or easier.
• Why? There are many more patterns that subjects
  minds can catch. If even one of the variations
  (optional, repeated, moved phrases) is helpful,
  three of these will be even more helpful. This
  is reflected in the transitional probabilities,
  which are much lower across phrases than within
  phrases.
• Prediction Test subjects do much better than
  control subjects.

47
Artificial Language Categorization
Test subjects do about as well as control
subjects for being able to categorize. This is
good, since it means subjects can abstract across
the artificial words and realize some belong to
the same category.
Day 5
Day 1
Day 5
48
Artificial Language Phrases
Test subjects much better than control subjects.
Second prediction is supported finding phrases
is easier when more variations are available,
even though there are more patterns to learn.
Day 5
Day 1
49
Statistically Learning Phrases

• Thompson Newport (2007) Adults can learn
  phrases in artificial languages if there are
  sentences that show the kinds of variation real
  sentences can have.
• Interesting Point When there are more variation
  types (optional, repeated, and moving phrases),
  adults are even better at unconsciously
  identifying phrases.
• Open Questions
• How well will this work for real language data?
  (Remember Gambell Yang (2006) found that
  transitional probabilities dont work so well for
  word segmentation when the data is realistic
  child-directed speech samples.)
• Will children be able use transitional
  probabilities to find phrases?

50
Questions?